Experience on Crowd-Human Challenge Zheng Ge1,2 Xin Huang1,2 Zequn Jie1,* Yuhu Shan1 1 Tencent AI Lab 2 Waseda University * Team leader

Baseline Re-implementation Exploring Techniques for Further Improvement

Baseline Re-implementation Baseline (visible body, paper): Faster-RCNN, Res50, FPN mMR: 55.94% mMR: 59.67% Baseline (visible body, ours): Faster-RCNN, Res50, FPN, ROI_align FPN+BN, Avoiding negative anchors in ignored regions mMR: 55.76%

Baseline Re-implementation ignore region anchor During training RPN, we avoid the negative anchors whose IoA (Intersection over Anchor) with an arbitrary labeled ignored region > 0.5. An example of labeled ignored region.

Baseline Re-implementation mMR AP@0.5 Baseline (paper) 55.94 85.60 Baseline (ours) 59.67 83.70 + Avoid negative anchors 58.53 84.92 + BN FPN 55.76 85.43 Table1. Evaluation results on Crowdhuman (visible body) validation set. Baseline (ours): Faster RCNN, Res50, FPN, ROI_align, nms_pre=6000

Baseline Re-implementation The effect of nms_pre. nms_pre mMR AP@0.5 12000 56.45 85.37 6000 55.76 85.43 2000 54.24 84.00 1500 54.00 83.38 1000 53.86 82.03 The best result on previous page Table2. Evaluation results on Crowdhuman (visible body) validation set.

Baseline Re-implementation Conclusion mMR AP@0.5 Baseline (paper) 55.94 85.60 Baseline (ours new) 54.24 84.00 Table3. Baseline comparison on validation set (visible body). mMR AP@0.5 Baseline (paper) 50.42 84.95 Baseline (ours new) 46.52 84.04 Table4. Baseline comparison on validation set (full body). Baseline (ours new): Faster RCNN, Res50, BN FPN, ROI_align, nms_pre=2000, Avoid negative anchors

Baseline Re-implementation Exploring Techniques for Further Improvement

Exploring Techniques Techniques bringing Techniques bringing large improvements Techniques bringing marginal improvements Techniques of limited effects Cascade R-CNN Deformable Conv Net SENet154 Multi-Scale Train/Test Ensemble SyncBN Focal Loss GIoU/IoU Loss COCO Pretrain Soft-NMS OHEM Adaptive-NMS Guided Anchor RPN Scale Balanced Sampling … Bounded Repulsion Loss R-CNN Context Merge Cityperson dataset

Exploring Techniques Conclusion mMR gain Baseline 46.52 - + Bounded RepGT Loss 45.80 0.72 + R-CNN Context 45.45 0.35 + multi-scale train/test 43.32 2.13 + Cascade R-CNN + Deformable Conv + SENet154 37.17 6.15 mMR gain Baseline 46.52 - + CityPerson 45.58 0.94 Baseline : Faster RCNN, Res50, BN FPN, ROI_align, nms_pre=2000, Avoid anchors

Exploring Techniques Final Ensemble for Submission SENet154 + Cascade R-CNN + DCN SENet154 + Cascade R-CNN + DCN + Bounded RepGt + Context SENet154 + Cascade R-CNN + DCN + Bounded RepGt + Context + cocopretrain SENet154 + Cascade R-CNN + DCN + Bounded RepGt + Context + cityperson SEResNeXt101 + Cascade R-CNN + DCN + Bounded RepGt + Context Coco没有增益，我们依旧使用它做ensemble的原因是希望coco能提供diversity, 使用senext101的理由也是如此

Exploring Techniques About Jaccard Index Score Threshold Thr. JI 0.55 75.78 0.50 76.48 0.45 77.01 0.40 77.46 Val. Set Val. Subset Test Set filtering huge gap best threshold: 0.55

Thanks